Recoleta
GitHub
Source note

Vlk: MemAct for the IDE – persistent working memory agents can prune themselves

Authors
akatsutki
Source
Hacker News
Published
Collection
Software Intelligence
Code IntelligenceIde AgentsPersistent MemoryMCP ServerAgent Context Management
Open GitHubMarkdown source

Summary

Vlk is a Rust MCP server that gives IDE coding agents persistent SQLite memory and a tool for deleting stale memory entries while saving a new lesson. It targets failure loops and context bloat in long-running coding-agent sessions.

Problem

  • Long-running IDE agents can keep retry errors, dead ends, and stale context, which wastes tokens and can trap the agent in loops.
  • Normal chat context often disappears after a restart, so coding agents lose useful lessons between sessions.
  • The problem matters for coding agents in Zed, Cursor, and Claude Desktop because they need durable task memory without carrying every failed attempt forward.

Approach

  • Vlk exposes one MCP tool, vlk_time_travel, over stdio JSON-RPC.
  • The agent calls the tool with target_mem_ids to remove and a learning string to keep.
  • vlk-core performs an atomic SQLite operation: delete selected rows from agent_history, then insert the lesson.
  • SQLite WAL stores the memory in vlk.db, so the IDE agent can reuse it after restarts.
  • The design follows MemAct from Zhang et al. 2025, where memory editing is an action the agent can choose at runtime.

Results

  • The excerpt gives no benchmark, user study, task success rate, latency result, or token-savings measurement.
  • It claims one exposed tool: vlk_time_travel.
  • In the shown example, memory slots 5, 6, and 7 are deleted and replaced with the lesson London API down, use cached 12°C.
  • The test command reports 1 slots pruned, N tokens saved, but the excerpt does not give a concrete value for N.
  • Supported clients named in the excerpt are Zed, Cursor, and Claude Desktop via MCP configuration.

Link